It seems to me that delaying self-driving cars for any appreciable amount of time will lead to more deaths/injuries, not fewer.
You're assuming that the quickest route to safest auto-driving lies via incremental improvements to existing systems and that's not necessarily true; see: AlphaGoZero
Fair point. I think a game with specific rules is probably better suited to the self-learning process that AlphaGo uses, whereas real-world driving rules are much more complex and nuanced. But I think it's still something worth considering.